← Downlink 2 Feb 2026 · CNBC

Orbital Data Centers: AI Leaves the Planet

2 Feb 2026 Source: CNBC Space · AI

The most unusual AI story of the season came from orbit: Nvidia-backed startup Starcloud trained and ran an AI model on a satellite, operating Google's Gemma on hardware already circling the planet. It is a proof of concept, but of something the industry has theorized about for years: data centers that never touch the ground.

The physics case is straightforward. Terrestrial data centers are constrained by three things: power, cooling and land. A large solar array in orbit captures constant energy, unhindered by night cycles, weather or clouds. Radiating heat into space is a thermal-engineering problem, but not an impossible one. And bandwidth to the ground keeps improving, which matters because the whole idea only works if the results come home.

The economic case is more interesting than the physics. Training frontier models is bottlenecked by electricity and by grid interconnect queues that now stretch for years. If even part of the training workload can move to platforms where energy is effectively free and continuous, the marginal cost of compute changes shape entirely. That is why investors are treating orbital compute as an energy play, not a rocket play.

The catch — and there is one — is latency. You will not run a chat session on a satellite; you will run batch workloads where a round trip of milliseconds-to-seconds is irrelevant: model training, simulation, and eventually the processing of satellite imagery at the source. That last one closes a neat loop: the satellites that collect the data become the machines that interpret it.

Running a small language model on a demo spacecraft is a long way from gigawatt-scale orbital compute. But so was one Falcon 1 from a reusable-booster landing program. The direction is what matters: compute is becoming a space industry, and space is becoming a compute platform.


Read the original →